Retinal vessel segmentation using textons

• Main Author: Zhang, Lei
• Published: University of East Anglia 2014
• Subjects: 004
• Online Access: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.665960

Segmenting vessels from retinal images, like segmentation in many other medical image domains, is a challenging task, as there is no unified way that can be adopted to extract the vessels accurately. However, it is the most critical stage in automatic assessment of various forms of diseases (e.g. Glaucoma, Age-related macular degeneration, diabetic retinopathy and cardiovascular diseases etc.). Our research aims to investigate retinal image segmentation approaches based on textons as they provide a compact description of texture that can be learnt from a training set. This thesis presents a brief review of those diseases and also includes their current situations, future trends and techniques used for their automatic diagnosis in routine clinical applications. The importance of retinal vessel segmentation is particularly emphasized in such applications. An extensive review of previous work on retinal vessel segmentation and salient texture analysis methods is presented. Five automatic retinal vessel segmentation methods are proposed in this thesis. The first method focuses on addressing the problem of removing pathological anomalies (Drusen, exudates) for retinal vessel segmentation, which have been identified by other researchers as a problem and a common source of error. The results show that the modified method shows some improvement compared to a previously published method. The second novel supervised segmentation method employs textons. We propose a new filter bank (MR11) that includes bar detectors for vascular feature extraction and other kernels to detect edges and photometric variations in the image. The k-means clustering algorithm is adopted for texton generation based on the vessel and non-vessel elements which are identified by ground truth. The third improved supervised method is developed based on the second one, in which textons are generated by k-means clustering and texton maps representing vessels are derived by back projecting pixel clusters onto hand labelled ground truth. A further step is implemented to ensure that the best combinations of textons are represented in the map and subsequently used to identify vessels in the test set. The experimental results on two benchmark datasets show that our proposed method performs well compared to other published work and the results of human experts. A further test of our system on an independent set of optical fundus images verified its consistent performance. The statistical analysis on experimental results also reveals that it is possible to train unified textons for retinal vessel segmentation. In the fourth method a novel scheme using Gabor filter bank for vessel feature extraction is proposed. The ii method is inspired by the human visual system. Machine learning is used to optimize the Gabor filter parameters. The experimental results demonstrate that our method significantly enhances the true positive rate while maintaining a level of specificity that is comparable with other approaches. Finally, we proposed a new unsupervised texton based retinal vessel segmentation method using derivative of SIFT and multi-scale Gabor filers. The lack of sufficient quantities of hand labelled ground truth and the high level of variability in ground truth labels amongst experts provides the motivation for this approach. The evaluation results reveal that our unsupervised segmentation method is comparable with the best other supervised methods and other best state of the art methods.